Creative Writing with an AI-Powered Writing Assistant: Perspectives from Professional Writers

Perhaps the most common desire was for a brainstorming partner; participants envisioned a tool that provided ideas a writer could riff off or helped with overcoming writer’s block. DH noted how brainstorming functionality would be especially useful for novice writers. Multiple participants described to us their process for working with human writin

It was generally agreed that a co-writing AI, like a human co-writer, is most useful when it complements a writer’s own skillset. If a writer already knows how to develop eloquent prose, the AI’s ability to suggest interesting ideas is more crucial than the exact language it chooses. Conversely, if a writer already has a clear story arc in mind, an

one hope participants had was for Wordcraft to be useful as a brainstorming and ideation tool. However, in practice, participants found it challenging to get suggestions that were interesting enough to be useful. RS questioned why he should bother using Wordcraft for it to simply “inject a few details I might have come up with on my own”. The need

Several participants wanted to be able to use Wordcraft to facilitate access to information. The idea of using the language model as a search engine repeatedly came up. In MT’s words: “I can’t read the entire internet or all the books by my favorite authors, but I can use a model like this one to leverage prolific catalogues” of information. DH hop

- Perhaps the most common desire was for a brainstorming partner; participants envisioned a tool that provided ideas a writer could riff off or helped with overcoming writer’s block. DH noted how brainstorming functionality would be especially useful for novice writers. Multiple participants described to us their process for working with human wri

One obvious step toward intentional language models would be for the creators of these models to be more intentional about what goes into their training data. For example, (Du et al., 2022) have demonstrated that a language model trained on a smaller dataset of curated data outperforms on standard benchmarks a model trained on a larger but less cur

A recurring theme in participant feedback was that the language model lacked taste and intentionality. It was capable of playing the “yes, and. . . ” improv game (Section 6.3), taking the user’s prompt as a given and running with it, but it lacked any narrative agenda of its own, which explains the abundance of clichés and generic tropes. In contra

professional creative writers are usually writing for a very particular audience, not the generic audience of the internet.